PHP2Shell probes: follow the destination beyond PHP

Deception Check Research · October 23, 2026 · All times UTC

The download address in a PHP-tagged request was already in our HTTP traversal, SSH, and telnet logs. The source of that request had submitted SSH delivery commands the previous evening. That is the useful connection in our PHP2Shell research: one web alert can give you a much wider hunt.

Start with the submitted destination, 217[.]60[.]102[.]5/sh, and the request source, 77[.]90[.]185[.]20. Search them as different roles across your web, SSH, and egress records, then check what the real service did next. Our decoys record what was sent; execution needs evidence from the actual service.

PHP2Shell refers to two PHP flaws: CVE-2026-109618 concerns corruption during out-of-memory handling; CVE-2026-109619 concerns stale state carried between requests handled by the same worker. An upload-progress marker alone does not establish either mechanism. PHP advisories.

The retained selection contains 26 unique requests from five source IPs between October 21 at 08:42:42 and October 22 at 21:55:25: 23 scanner/reconnaissance requests carrying PHP2Shell tool labels, one downloader input carrying a php2shell argument, and two separate WordPress inclusion attempts. Keeping that last pair separate matters: an upload-progress field is not a fingerprint for the new PHP flaws.

What we observed

Swipe tables to see all columns.

Observed category Requests What it supports
Self-labelled multipart probes 18 Deliberate PHP2Shell-related probing
Self-labelled probes, body unavailable 3 Request-path and tool-label evidence
Self-labelled GET reconnaissance 2 A tool identifying itself with PHP2Shell
Extensive VC upload/session-file inclusion attempts 2 Separate WordPress plugin technique; not PHP2Shell
php2shell-tagged downloader input 1 Session-file inclusion attempt with an attacker-supplied vector label

The event index preserves all 26 records, their selection rules, timestamps, and revised classifications. The WordPress pair remains available for comparison, not folded into a PHP2Shell exploit count.

26 SELECTED REQUESTS / THREE DISTINCT CLASSES

PHP2Shell tool-labelled probes / recon23
Sender-tagged downloader input1
Separate Extensive VC inclusion inputs2
23 tool-labelled requests from three sources, one sender-tagged downloader input from a fourth, and two separate WordPress attempts from a fifth.

Before which disclosure?

21 OCTOBER / UTC

08:42:42Our first selected probe
13:34GitHub advisories published
14:00Scheduled release target
Public leaks preceded the probe. Milestones are not drawn to an elapsed-time scale. Read the full sourced timeline.

Our first selected probe arrived at 08:42:42 on October 21, about 4 hours 51 minutes before GitHub published PHP's two advisories at 13:34:13 and 13:34:25. It also preceded the scheduled 14:00 release. That is early relative to the advisories, but vulnerability information was already public.

The PHP Foundation describes circulating CVE identifiers and an accidental demonstration video on October 20, followed by an accelerated release. The full timeline carries the wider disclosure and distributor chronology, including differing accounts of early package availability. PHP Foundation account.

Other defenders also saw pre-advisory activity. AWS describes an October 17 attempt and further activity from around 02:00 on October 21. ELLIO reports its first labelled request at 08:36. VulnCheck's 08:37 to 08:52 observations include 85[.]11[.]167[.]243, the same address as our 08:42:42 probe. AWS MadPot, ELLIO, VulnCheck.

RESEARCH WORKBENCH

The full timeline, in one place

Follow discovery, private reporting, public disclosure, and sensor observations together. Each event retains its own source and time precision. This is an ordered ledger, not an elapsed-time chart or a claim that one event caused another.

Showing all 27 events.

  1. 2024 (exact date unspecified)UTC where timed

    Discovery and disclosure

    First signs of the OOM problem found

    Theoretical exploitation concern then, not a demonstrated RCE vector left knowingly unfixed.

    Source · year, Foundation account · discovery account

  2. 2026-08-11UTC where timed

    Discovery and disclosure

    Calif reports initial RCE

    Private vulnerability reporting, not public exploit publication.

    Source · day, researcher account · third-party account

  3. 2026-08-14UTC where timed

    Discovery and disclosure

    Calif supplies private RCE proof of concept

    PHP confirmation reported August 15.

    Source · day, researcher account · third-party account

  4. 2026-10-17UTC where timed

    Other sensor reports

    AWS reports an exploitation attempt

    AWS leaves real-world success and exploit acquisition unclear.

    Source · day, AWS telemetry report · third-party telemetry

  5. 2026-10-20 03:38:56.597539UTC where timed

    Our observations

    Shared script destination in earlier HTTP input

    HTTP traversal input names the same destination later submitted in our October 22 sender-tagged PHP request. The connection is the destination, not evidence of PHP2Shell use on October 20.

    Source · microsecond string, sensor-recorded · first-party shared-delivery observation

  6. 2026-10-20 ~15:00UTC where timed

    Discovery and disclosure

    CVE identifiers circulate

    Initial leak described as nontechnical.

    Source · approximate, Foundation account · disclosure account

  7. 2026-10-20 19:30UTC where timed

    Discovery and disclosure

    PHP warning; release advanced by one day

    New target October 21, 14:00; original October 22, 14:00.

    Source · minute, Foundation account · disclosure account

  8. 2026-10-20 ~20:00UTC where timed

    Discovery and disclosure

    Accidental demonstration video

    Limited technical details; copies reportedly circulated.

    Source · approximate, Foundation account · disclosure account

  9. 2026-10-20 ~23:00UTC where timed

    Discovery and disclosure

    Mitigation guidance shared privately

    Possible further leakage, not established.

    Source · approximate, Foundation account · disclosure account

  10. 2026-10-21 ~02:00 to 08:00UTC where timed

    Other sensor reports

    AWS reports PoCs and more attempts

    None of its six published source IPs matches our selected five.

    Source · approximate, AWS account · third-party telemetry

  11. 2026-10-21 06:07:42UTC where timed

    Discovery and disclosure

    Public comment reports patched-image availability

    Dates the statement, not first registry publication or image contents.

    Source · second, HN platform metadata · timestamped statement

  12. 2026-10-21 06:17:04.639046UTC where timed

    Our observations

    Linked delivery endpoint appears over SSH

    This SSH input names the same endpoint later used in the October 22 sender-tagged PHP request. The time marks the first retained vector-bearing curl fragment; the preceding host-bearing SCP fragment is 06:17:04.630082. Across the selected evidence window: 21 loader sessions from 2 sources, not 21 proven executions.

    Source · microsecond string, sensor-recorded · first-party shared-delivery observation

  13. 2026-10-21 06:18:17UTC where timed

    Discovery and disclosure

    Chainguard CEO acknowledges early release

    Cites exploitation reports; underlying evidence not supplied there.

    Source · second, HN platform metadata · first-party statement

  14. 2026-10-21 07:24:56.429853UTC where timed

    Our observations

    Linked delivery endpoint appears over telnet

    This telnet input names the same endpoint later used in the October 22 sender-tagged PHP request, extending the linked activity across delivery methods. Across the selected evidence window: 16 loader sessions from 15 sources. Parser-split input is not proof of execution. HTTP requests and SSH/telnet sessions are different units and must not be summed as attacks or infections.

    Source · microsecond string, sensor-recorded · first-party shared-delivery observation

  15. 2026-10-21 07:41UTC where timed

    Discovery and disclosure

    Foundation's time for early Chainguard images

    Earlier statements predate this. First registry publication remains unverified.

    Source · minute, Foundation account · disputed first-availability chronology

  16. 2026-10-21 08:36UTC where timed

    Other sensor reports

    ELLIO reports its first labelled request

    Self-labelled PHP2Shell request. No exact source-IP overlap with our selected cohort established here.

    Source · minute, ELLIO telemetry report · third-party telemetry

  17. 2026-10-21 08:37 to 08:52UTC where timed

    Other sensor reports

    VulnCheck sees 85.11.167.243 probes

    Exact source and early user-agent overlap with our selection.

    Source · minute, VulnCheck telemetry report · third-party telemetry

  18. 2026-10-21 08:42:42.137613UTC where timed

    Our observations

    Our first selected labelled upload probe

    85.11.167.243; body retained as capture placeholder, not exploit validation.

    Source · microsecond string, sensor-recorded · first-party observation

  19. 2026-10-21 13:31UTC where timed

    Discovery and disclosure

    Public release tarballs pushed

    Push chronology attributed to Daniel Scherzer; activity views not independently retrieved.

    Source · minute, release-manager account · release-manager account

  20. 2026-10-21 13:33UTC where timed

    Discovery and disclosure

    Website announcement repository pushed

    Not an exact independent measurement of php.net deployment.

    Source · minute, release-manager account · release-manager account

  21. 2026-10-21 13:34:13UTC where timed

    Discovery and disclosure

    Stale-state advisory published

    4h51m30.862387s after our first selected request.

    Source · second, GitHub published_at · platform publication metadata

  22. 2026-10-21 13:34:25UTC where timed

    Discovery and disclosure

    OOM advisory published

    4h51m42.862387s after our first selected request.

    Source · second, GitHub published_at · platform publication metadata

  23. 2026-10-21 13:35UTC where timed

    Discovery and disclosure

    Fixed and backport branches pushed

    Git object author/tag dates are not first-publication evidence.

    Source · minute, release-manager account · release-manager account

  24. 2026-10-21 14:00UTC where timed

    Discovery and disclosure

    Advanced coordinated release target

    Do not substitute this deadline for exact advisory or package publication.

    Source · scheduled deadline · planned event

  25. 2026-10-21 14:15 to 14:24UTC where timed

    Other sensor reports

    VulnCheck sees 45.198.224.169 probes

    Our selected request from this address is the next day, not this same event.

    Source · minute, VulnCheck account · third-party telemetry

  26. 2026-10-22 21:53:19.178708 and 21:54:17.455252UTC where timed

    Our observations

    Our two filename-injection uploads

    Separate Extensive VC inclusion pattern using options[template], documented as CVE-2023-0159. Not PHP2Shell.

    Source · sensor-recorded · first-party observation

  27. 2026-10-22 21:55:25.787307UTC where timed

    Our observations

    First retained PHP-tagged downloader request

    Same delivery host/path as earlier HTTP, SSH, and telnet inputs. Source 77.90.185.20 also appears in SSH. This dates our retained observation, not the first adoption of the vector. The delivery pattern is RedTail-consistent; the script was not retrieved.

    Source · sensor-recorded · first-party observation

Timeline JSON · Full-resolution timeline image

How the later request connects to wider delivery activity

OBSERVED INPUTS / SHARED DESTINATION

HTTP traversal104 requests
SSH21 sessions
Telnet16 sessions
PHP1 downloader request
SAME HOST + PATH217[.]60[.]102[.]5/shNamed in submitted commands
Four distinct input routes, not four stages of execution. The PHP source also appears in SSH.

At 21:55:25 on October 22, a request from 77.90.185.20 placed PHP code in an upload-progress field. If executed, it would fetch hxxps://217[.]60[.]102[.]5/sh with wget or curl, then pipe the result to a Unix shell with a php2shell argument. A following request referenced a session-file path associated with the submitted session identifier.

That is a session-file inclusion attempt. php2shell is the sender's label passed to the shell invocation, not our validation of a PHP2Shell exploit. The shared destination and earlier SSH input remain useful regardless of that label.

The broader retained logs contain other inputs requesting the same HTTPS host and /sh path, with apache, ssh, or telnet arguments. Those arguments are visible in the submitted commands; we have not inspected how the remote script uses them.

Swipe tables to see all columns.

Delivery route Retained evidence Source addresses
Apache-style HTTP traversal 104 distinct requests 47
SSH 21 delivery sessions 2
Telnet 16 delivery sessions 15
php2shell-tagged session-file attempt 1 request 1

Requests and sessions are different counting units. The PHP request is already included in the 26-request selection; the table is a separate view of shared delivery activity, not an additional attack total. Counts and method.

The HTTP delivery pattern appears in our retained evidence by 03:38:56 on October 20. There is a second connection beyond the destination: 77.90.185.20 submitted the SSH route at 21:25:51 on October 21, then the PHP route the following evening. That combination makes this a useful cross-protocol hunt.

All 104 HTTP traversal requests use libredtail-http; the PHP downloader request has no user agent. Published SANS observations and Akamai's sample-based RedTail research describe comparable traversal, downloader, and multi-exploit delivery patterns. Our assessment is RedTail-consistent delivery, not confirmed malware-family identification: we did not fetch the remote script or obtain an installed sample. SANS observations, Akamai's RedTail analysis.

URLhaus also records an October 4 malware-download URL on the same destination address at /clean. That is historical context for the host, not validation of our /sh path. URLhaus record.

What the captured uploads tell us

Padding is part of the request, not a malware file

The earliest requests used php2shell-scan/109618, targeted /index.php, and declared 62,935-byte bodies. The retained body fields contain capture placeholders. Associated capture records identify uploads named a and b.

Static inspection of three retained copies of each file reproduced their logged hashes and established their complete contents:

Swipe tables to see all columns.

File Bytes Complete retained content
a 10,370 CRLF, then 5,184 A bytes and 5,184 B bytes
b 64 64 C bytes

VulnCheck describes the same a/b upload pattern as an attempt to cause memory exhaustion, and reports that its observed requests did not exhaust memory in a single request. Our retained files are padding, consistent with that interpretation. For an OOM investigation, multipart structure, allocation behavior, and worker errors matter more than finding executable content inside each upload. The byte-analysis record contains the offsets and hashes; use them for artifact correlation. VulnCheck's observations.

Those files do not reconstruct the full multipart request. Three later records retain only 4,096 characters of declared 62,929-byte bodies; three other labelled probes retain no body. That capture gap matters when assessing the conditions required to trigger either PHP vulnerability.

A separate WordPress technique caught by the same selection

Two October 22 uploads from 45.115.26.26 placed PHP in an uploaded filename, while the visible file content was Z. The code prints EXTVC_OK and attempts to pass a query-string value to system().

The following requests use extensive_vc_init_shortcode_pagination and options[template], matching the Extensive VC Addons local-file inclusion pattern documented for CVE-2023-0159. The upload-progress field supplies a session-file poisoning technique; it does not make this PHP2Shell. They also put c=id in POST data while the injected code reads $_GET['c'], an observable mismatch. The replay retains that distinction. WPScan advisory, Exploit-DB 52085.

RESEARCH WORKBENCH

Walk through the retained evidence

These walkthroughs reconstruct the logs, not exploitation on a real PHP server. Playback uses one record every four seconds, not the original timing; no code or network request is executed. All steps are also written out below each replay.

01

Request received

Incoming method, path, source, and retained input.

02

Artifact inspected

Retained bytes checked against recorded hashes.

03

Response recorded

Status and generation metadata describe the decoy.

04

Runtime evidence needed

Actual worker, process, file, and egress records.

Padding files: investigate the request and worker

All three retained copies of each artifact match. These repetitive bytes are padding. In an OOM investigation, inspect multipart structure and worker memory behavior; these files alone do not reconstruct the full request.

a · 10,370 bytes
25,184 × A5,184 × B
b · 64 bytes
64 × C

Exact sequence, not a shared length scale. File a starts with CRLF; its two large runs are equal in length. File b is shown on its own track.

Offsets, hashes, and static analysis JSON

Separate Extensive VC filename injection and inclusion attempts

Sanitized reconstruction from retained records.

Step 1 of 4

Upload request

RECORDED INPUT / CONTEXT

POST /wp-login.php; multipart filename contains PHP that prints EXTVC_OK and reads a command from the GET parameter c. Session identifier redacted.

RECORDED RESPONSE / ARTIFACT

Recorded HTTP 200; response_length 1913; generated response flag true.

RUNTIME QUESTION

No verified PHP execution or session-file write.

Full sequence transcript

  1. POST /wp-login.php; multipart filename contains PHP that prints EXTVC_OK and reads a command from the GET parameter c. Session identifier redacted.

    Recorded HTTP 200; response_length 1913; generated response flag true.

    No verified PHP execution or session-file write.

  2. POST /wp-admin/admin-ajax.php; form field options[template]=/tmp/sess_[redacted]; c=id is in the POST body; logged query is empty.

    Recorded HTTP 404; response_length 1259; generated response flag true.

    No verified session-file inclusion or command execution.

  3. POST /wp-admin/admin-ajax.php; form field options[template]=/var/lib/php/sessions/sess_[redacted]; c=id is in the POST body; logged query is empty.

    Recorded HTTP 404; response_length 923; generated response flag true.

    No verified session-file inclusion or command execution.

  4. POST /wp-admin/admin-ajax.php; form field options[template]=/var/lib/php/session/sess_[redacted]; c=id is in the POST body; logged query is empty.

    Recorded HTTP 404; response_length 1259; generated response flag true.

    No verified session-file inclusion or command execution. The submitted filename reads GET c, while these requests place c in POST.

Retained evidence, October 22, 2026; observed source 45[.]115[.]26[.]26; first of two comparable target sequences

Sender-tagged downloader input and a session-path request

Sanitized reconstruction from retained records.

Step 1 of 2

Upload-progress request

RECORDED INPUT / CONTEXT

POST /index.php; retained body prefix places PHP inside PHP_SESSION_UPLOAD_PROGRESS. It requests retrieval from hxxps://217[.]60[.]102[.]5/sh and pipes retrieved content to a shell with the argument php2shell. Session identifier redacted.

RECORDED RESPONSE / ARTIFACT

Recorded HTTP 404; response_length 0; generated response flag true.

RUNTIME QUESTION

No verified PHP execution, outbound retrieval, or shell execution.

Full sequence transcript

  1. POST /index.php; retained body prefix places PHP inside PHP_SESSION_UPLOAD_PROGRESS. It requests retrieval from hxxps://217[.]60[.]102[.]5/sh and pipes retrieved content to a shell with the argument php2shell. Session identifier redacted.

    Recorded HTTP 404; response_length 0; generated response flag true.

    No verified PHP execution, outbound retrieval, or shell execution.

  2. GET /index.php?page=/tmp/sess_[redacted]; the raw session-path value matches the preceding POST cookie. Cookie and session identifier redacted.

    Recorded HTTP 200; response_length 0; generated response flag true.

    No verified session-file inclusion or payload result. Two later GETs use different session identifiers and are outside this selected two-step replay.

Retained evidence, October 22, 2026; observed source 77[.]90[.]185[.]20; selected POST and immediately following GET

Read the selection and counting method

Selection and counting method

Selection uses PHP2Shell identifiers or upload-progress markers in incoming request fields, excluding generated responses. Twenty-eight matching rows reduce to 26 unique original request lines after exact-hash deduplication. Eleven contain a source address; 15 use a connection record with the same sensor identity and session. No timing-only or cross-sensor joins are used.

Classify after selecting: 23 tool-labelled probes/reconnaissance requests from three addresses; one php2shell-tagged downloader input from a fourth; and two separate Extensive VC uploads from a fifth. Three labelled requests have no retained bodies and stay separate from the multipart category. The public event index preserves stable IDs and original-line hashes so readers can trace the counts; archive custody and review records are retained privately.

Source joins require the same sensor identity and session. A matching session identifier is not proof that two requests reached the same PHP worker.

26-request event index · Sanitized replay data · Standalone evidence guide

Scanner labels are context, not identity

VulnCheck independently reports two addresses in our selection: 85.11.167.243 and 45.198.224.169. The first overlaps our early probe interval; our request from the second is on the following day.

ELLIO reports Mirai-loader delivery from 45[.]198[.]224[.]169 on October 21. That is useful additional behavior from another sensor network, not a malware-family finding for our later request. Dated ASN, registry, and Shodan context remains alongside the IOC table. ELLIO observations.

Treat user agents the same way. shellflow-php2shell/1.0 appears in a public tool's source, but a copied header does not identify who ran it. Public tool source.

RESEARCH WORKBENCH

Indicators and network context

Use the addresses below to find related records, then follow the activity. The five source addresses and the submitted download destination have different roles. This is an investigation list, not a blanket blocklist.

ASN and location describe provider records, not an operator or an attacker’s location. Country labels explicitly distinguish GeoIP estimates from registry information. Lookup context is dated; it is not a reconstruction of routing at the moment of each request.

Request windows shown to whole seconds in UTC; full stored timestamps are in the event index. On narrow screens, each network row becomes a labelled card.

Address and observed roleSelected requests / first to last UTCASN / networkLocation contextSource and interpretation
85[.]11[.]167[.]243observed request source20
AS197170TechTies Inc.; 85.11.167.0/24BulgariaShodan GeoIPCurrent AS197170 differs from AS209630 in VulnCheck’s October 21 report. Historical routing remains unresolved.Record dates: ASN/prefix: 2026-10-23; retained Shodan view last seen: 2026-10-21RIPEstat route · RIPEstat ASN · Shodan host data
185[.]28[.]119[.]18observed request source2
AS62005BlueVPS OU; 185.28.119.0/24United StatesShodan GeoIPShodan data also carries an open-dir tag, a service observation rather than an intent assessment.Record dates: ASN/prefix: 2026-10-23; retained Shodan view last seen: 2026-10-21RIPEstat route · RIPEstat ASN · Shodan host data
45[.]198[.]224[.]169observed request source1
AS215925VPSVAULT.HOST LTD; 45.198.224.0/24CanadaAFRINIC registration country; GeoIP not verifiedShodan scanner tag; an explicit GeoIP country was not verified.Record dates: 2026-10-23RIPEstat route · RIPEstat ASN · Registry record · Shodan host data
45[.]115[.]26[.]26observed request source2
AS199457SolidCore Hosting LTD; 45.115.26.0/24VietnamAPNIC registration country; GeoIP not verifiedAPNIC assignment is a /26 inside the currently announced /24. Registry and routing records describe different layers.Record dates: 2026-10-23RIPEstat route · RIPEstat ASN · Registry record
77[.]90[.]185[.]20observed request source1
AS213790Limited Network LTD; 77.90.185.0/24LithuaniaShodan GeoIPShodan scanner tag. This source also appears in the retained SSH delivery activity.Record dates: ASN/prefix: 2026-10-23; retained Shodan view last seen: 2026-10-23RIPEstat route · RIPEstat ASN · Shodan host data
217[.]60[.]102[.]5submitted downloader destinationNot a source count
AS209373SWISSNET LLC; 217.60.102.0/24NetherlandsRIPE registration country; GeoIP not verifiedSubmitted retrieval destination. These records do not establish a callback to it.Record dates: 2026-10-23RIPEstat route · RIPEstat ASN · Registry record

Tool strings, destination, and captured-file hashes

Swipe or focus the reference tables to scroll horizontally.

ValueRoleHow to use it
hxxps://217[.]60[.]102[.]5/shsubmitted retrieval URLDo not browse or fetch. /sh bytes not retrieved or analyzed.
libredtail-httpcaller-supplied user agent in 104 distinct HTTP traversal requestsRedTail-consistent published indicator, not a standalone family or operator verdict. The php2shell-tagged downloader input has no user agent.
150ae88f0999032e2a6113501d6082f54a3ed98d02427b21a96e773a27d7bf16byte-verified filler upload-artifact SHA256Exact repetitive bytes verified; artifact correlation, not malware classification, operator identity, or PHP2Shell trigger/success proof.
27d926a6708d6769d97d065e7a5a860ab496eab272c03d1cebc55ec71b954fa2byte-verified filler upload-artifact SHA256Exact repetitive bytes verified; artifact correlation, not malware classification, operator identity, or PHP2Shell trigger/success proof.
php2shell-scan/109618caller-supplied user agentMutable tool label, not ownership or CVE validation.
php2shell-109618caller-supplied user agentMutable tool label, not ownership or CVE validation.
shellflow-php2shell/1.0caller-supplied user agentMutable tool label, not ownership or CVE validation.
Additional request markers and contextual strings

These broad strings need surrounding structure and activity. Do not alert on generic field names, paths, or shell-tool names alone.

ValueRoleHow to use it
apacheargument submitted to the shared downloader/shell invocationIncoming vector label, not a verified infection tag or evidence of executed loader behavior.
sshargument submitted to the shared downloader/shell invocationIncoming vector label, not a verified infection tag or evidence of executed loader behavior.
telnetargument submitted to the shared downloader/shell invocationIncoming vector label, not a verified infection tag or evidence of executed loader behavior.
php2shellargument submitted to the shared downloader/shell invocationIncoming vector label, not a verified infection tag or evidence of executed loader behavior.
dlrremote transfer account name submitted in SSH command textAttacker-supplied retrieval context only. No credential or key material distributed; no account accessed.
B7kQpoc618selected-request multipart boundaryCorrelation marker; easy to copy; not actor identity.
EXTINJselected-request multipart boundaryCorrelation marker; easy to copy; not actor identity.
WebKitFormBoundaryn407ilexselected-request multipart boundaryCorrelation marker; easy to copy; not actor identity.
EXTVC_OKoutput marker present in submitted PHPNot observed response output or execution proof.
__START__output marker present in submitted PHPNot observed response output or execution proof.
__END__output marker present in submitted PHPNot observed response output or execution proof.
PHP_SESSION_UPLOAD_PROGRESScontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
PHPSESSIDcontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
fcontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
acontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
bcontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
php2shellcontextual input name, filename, or argumentGeneric or caller-controlled; not a standalone malicious signature.
/requested pathPath intent, not installed software or successful execution.
/index.phprequested pathPath intent, not installed software or successful execution.
/lib/phpunit/phpunit/Util/PHP/eval-stdin.phprequested pathPath intent, not installed software or successful execution.
/wp-login.phprequested pathPath intent, not installed software or successful execution.
/wp-admin/admin-post.phprequested pathPath intent, not installed software or successful execution.
/wp-admin/admin-ajax.phprequested pathPath intent, not installed software or successful execution.
extensive_vc_init_shortcode_paginationcontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
options[template]contextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
local-downloadcontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
local-destination-id=0contextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
../../../wp-config.phpcontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
/etc/passwdcontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
c=idcontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
pagecontextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
/tmp/sess_[redacted]contextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
/var/lib/php/sessions/sess_[redacted]contextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
/var/lib/php/session/sess_[redacted]contextual inclusion/file-read sequence featureNeighbor records, not necessarily selected-cohort members; no actual file read or inclusion established.
system()function/tool named in incoming payloadNot actual process telemetry. Complete executable payload deliberately omitted.
$_GET['c']function/tool named in incoming payloadNot actual process telemetry. Complete executable payload deliberately omitted.
wgetfunction/tool named in incoming payloadNot actual process telemetry. Complete executable payload deliberately omitted.
curlfunction/tool named in incoming payloadNot actual process telemetry. Complete executable payload deliberately omitted.
shfunction/tool named in incoming payloadNot actual process telemetry. Complete executable payload deliberately omitted.

Indicator JSON · Plain-text indicators · Dated network-context records

What to do with this evidence

Patch, then verify the running deployment. PHP's first fixed supported-branch releases are 8.2.35, 8.3.36, 8.4.27, and 8.5.12. Supplier backports may use different package versions. Confirm the deployed build and the runtime loaded by workers after restart or replacement; a downloaded package or banner is not enough. PHP advisory, Chainguard package guidance.

October 23 update: PHP separately announced CVE-2026-115560, an unlikely session-upload-progress RCE vector under a specific combination of application code and configuration. PHP suggests setting session.upload_progress.enabled=0 until the next regular bugfix release; applications that never change the session serialize handler at runtime can keep the feature enabled. Follow the new PHP notice as well as updating for the two PHP2Shell flaws.

Follow the request into other telemetry. Start with the role-labelled indicators, then correlate upload structure, PHP-bearing filenames, session-path requests, worker failures, unexpected child processes, file changes, and outbound connections. Search the destination and source across HTTP, SSH, and telnet records where available. For stale-state analysis, establish the actual PHP worker across requests, not just a matching session identifier.

Cover the service you actually run. Include Apache httpd/apache2 when PHP is loaded as a module, not only php-fpm or php-cgi. Hunt the destination in outbound telemetry, libredtail-http in HTTP logs, and downloader/shell command lines in endpoint records. Worker crashes and memory-exhaustion bursts are triage leads to correlate with requests, not stand-alone exploit verdicts. The on-page detection appendix includes these separate paths.

Record what the sensor retained and what it generated. The later web-decoy sequences recorded generated responses. A 200, 404, capture label, or handler counter cannot settle whether a real PHP worker executed code. Keep declared body length, retained prefix length, capture reference, and response provenance as separate fields.

Use the technical evidence guide, event index, and sourced timeline to structure the investigation. The useful next question is not only “did this address hit us?” It is “what did the running service do next?”

RESEARCH WORKBENCH

Research detection appendix

Start with the telemetry you have. Request-pattern rules find the input seen in this investigation; endpoint hunts look for what a real service did next. Neither is a universal exploit detector. Each entry states its prerequisites, tuning needs, and the checks actually performed.

WEB / WAF / IDS

Tool-labelled requests and suspicious upload structure. Encrypted HTTP needs visibility after TLS termination.

ENDPOINT / SIEM

PHP worker process ancestry, command lines, network activity, and file changes.

OFFLINE CONTENT

YARA triage of captured request text, not proof that a payload ran or that filler is malware.

YARA compiled and passed inert fixtures. Sigma parsed and passed local predicate checks. Snort and platform queries received static syntax/schema review; their engines and tenants were not exercised. Supplemental local checks cover destination/direction predicates and error text, not platform execution. Download both fixture files and the validator with the rules.

Download the complete detection pack
Deployment and tuning guide · Inert test fixtures · Local validator

yaml

Sigma: PHP2Shell labels in HTTP

Find requests whose caller-supplied User-Agent names PHP2Shell.

Telemetry
Parsed web-server or reverse-proxy request logs. Map the incoming User-Agent value to Sigma cs-user-agent; verify category webserver in the conversion pipeline.
What it looks for
Case-insensitive substring php2shell covers the three observed labels. This is a probe/research lead and can also match authorized testing.
Validation
Parsed with pySigma 2.0.1. This rule passed seven inert local predicate fixtures. Backend conversion, target-engine execution, and alert deployment were not performed.
Tuning
Authorized scanners, copied headers, research, and replayed logs.
Coverage limits
Missing or changed User-Agent, including the observed downloader request with no UA. A match does not establish exploitation, execution, or operator identity.
Read yaml source
title: PHP2Shell Label in HTTP User Agent
id: 6e9d7efe-1b45-46dd-9b17-3accb7b8654e
status: experimental
description: >-
    Finds caller-supplied PHP2Shell labels, including the three labels in this
    research. A match is a research or probe lead, not a validated exploit,
    malicious operator identity, or evidence of execution.
author: Deception Check
date: '2026-10-23'
references:
    - https://sigmahq.io/docs/basics/log-sources.html
logsource:
    category: webserver
    definition: 'Map cs-user-agent to the parsed incoming HTTP User-Agent value.'
detection:
    selection:
        cs-user-agent|contains: 'php2shell'
    condition: selection
fields:
    - c-ip
    - cs-method
    - cs-uri-query
    - cs-user-agent
falsepositives:
    - Authorized testing, research scanners, copied labels, and log replay
level: low

Download sigma-web-php2shell-label.yml

Documentation: Sigma field taxonomy · Sigma rule specification

yaml

Sigma: Linux PHP worker starts a shell or downloader

Find generic runtime behavior that deserves investigation after a PHP request.

Telemetry
Linux process-creation events with full Image and direct ParentImage paths; preserve process identifiers, command lines, and host context.
What it looks for
A php-fpm or php-cgi basename, optionally followed by numeric version components, directly parents sh, bash, dash, ksh, zsh, curl, or wget. The regexes match executable-path fields rather than request text.
Validation
Parsed with pySigma 2.0.1; 13 inert local predicate fixtures passed. No Sigma backend conversion or production telemetry execution.
Tuning
Applications and maintenance tasks intentionally launching these tools. Baseline per service and investigate command lines.
Coverage limits
Generic behavior, not CVE-specific. Misses PHP CLI, Apache mod_php, renamed workers/tools, wrappers, busybox, grandchildren, absent ancestry, and execution without these children.
Read yaml source
title: Linux PHP Worker Starts Shell or Downloader
id: 0fc952b5-d91d-492c-93a2-c091588cf23a
status: experimental
description: >-
    Finds a direct shell or downloader child of a php-fpm or php-cgi process.
    This is a generic runtime investigation lead, not a PHP2Shell signature.
    Add verified worker names used in your deployment and baseline expected jobs.
author: Deception Check
date: '2026-10-23'
references:
    - https://sigmahq.io/docs/basics/log-sources.html
logsource:
    category: process_creation
    product: linux
    definition: 'Requires full executable paths for Image and direct ParentImage.'
detection:
    selection_parent:
        ParentImage|re: '/php-(fpm|cgi)([0-9]+([.][0-9]+)*)?$'
    selection_child:
        Image|re: '/(sh|bash|dash|ksh|zsh|curl|wget)$'
    condition: selection_parent and selection_child
fields:
    - Image
    - CommandLine
    - ParentImage
    - ParentCommandLine
    - User
    - ProcessId
    - ParentProcessId
falsepositives:
    - Application features that intentionally invoke shells, downloaders, or maintenance tasks
level: medium

Download sigma-linux-php-worker-child.yml

Documentation: Sigma field taxonomy · Sigma log-source guidance

yara

YARA: offline input triage and exact filler correlation

Triage locally retained text or decoded HTTP requests; keep filler correlation separate from suspicious-input signals.

Telemetry
Offline files, ideally one decoded request/body per file. Text rules inspect ASCII tokens and cap input at 2 MiB. The exact-artifact rule uses file size and SHA-256.
What it looks for
Three roles: weak php2shell text label; stronger co-occurrence of upload-progress, a PHP opening tag, and system/shell_exec/passthru function tokens; exact correlation of the two verified repetitive filler files. Co-occurrence does not bind code to a multipart field.
Validation
Compiled without warnings using yara-python 4.5.4. Fourteen inert match fixtures passed, including reconstructed filler hashes, one-byte mutations, ordinary repeated filler, and the size guard. This is offline fixture validation, not exploit testing.
Tuning
Research documents, source-code uploads, and training examples can match text rules. Exact filler bytes may be reused by unrelated tests.
Coverage limits
No PHP parser or decoding of JSON escapes, HTML entities, archives, or encryption. Concatenated logs can combine unrelated tokens. Filler matches are not malware classifications and do not establish surrounding-request safety.
Read yara source
// Offline files only. Supply one decoded request/body per file where possible.
// Do not use these rules to classify the two filler artifacts as malware.
import "hash"

rule DC_PHP2Shell_Text_Label : research_label low_confidence
{
    meta:
        author = "Deception Check"
        date = "2026-10-23"
        purpose = "Weak text correlation; documentation and copied labels also match"
        scope = "Offline UTF-8/ASCII text, at most 2 MiB; not a CVE or malware signature"
    strings:
        $label = "php2shell" ascii nocase
    condition:
        filesize <= 2MB and $label
}

rule DC_Upload_Progress_PHP_Function_Tokens : input_triage stronger_context
{
    meta:
        author = "Deception Check"
        date = "2026-10-23"
        purpose = "Co-occurring upload-progress, PHP opening tag, and command-function tokens"
        scope = "Stronger than a label alone; no parsing, field binding, execution, or CVE claim"
    strings:
        $progress = "PHP_SESSION_UPLOAD_PROGRESS" ascii
        $php = "<?php" ascii nocase
        $call = /\b(system|shell_exec|passthru)\s*\(/ ascii nocase
    condition:
        filesize <= 2MB and all of them
}

rule DC_Known_Repetitive_Upload_Artifact : filler_correlation not_malware_classification
{
    meta:
        author = "Deception Check"
        date = "2026-10-23"
        purpose = "Exact byte correlation with the two verified repetitive filler uploads"
        scope = "A matching artifact is filler; surrounding request safety remains unresolved"
    condition:
        (filesize == 10370 and hash.sha256(0, filesize) ==
            "150ae88f0999032e2a6113501d6082f54a3ed98d02427b21a96e773a27d7bf16") or
        (filesize == 64 and hash.sha256(0, filesize) ==
            "27d926a6708d6769d97d065e7a5a860ab496eab272c03d1cebc55ec71b954fa2")
}

Download php2shell-offline-triage.yar

Documentation: YARA rule syntax · YARA hash module

snort

Snort 3: visible HTTP labels and body-token combinations

Find a PHP2Shell User-Agent label or a multipart request containing upload-progress/PHP/command-function tokens.

Telemetry
Snort 3 HTTP inspection over plaintext HTTP or an authorized decrypted feed. Ordinary encrypted HTTPS does not expose these buffers.
What it looks for
SID 1000618 searches the User-Agent sticky buffer. SID 1000619 requires multipart Content-Type and co-occurring tokens in http_client_body. It does not parse the multipart field relationship or establish execution.
Validation
Official Snort 3 header, service, body, content, and PCRE syntax reviewed; four inert local body-predicate fixtures passed. Snort compilation, PCAP replay, HTTP normalization, and stream/depth behavior were not tested.
Tuning
Authorized probes, copied labels, and source/training uploads containing the combined tokens.
Coverage limits
TLS visibility, HTTP inspector and request-body depth, truncation, encodings, changed labels, and altered code tokens. Check local SID collisions and HOME_NET. Keep as an investigation alert until locally validated.
Read snort source
# Snort 3, experimental hunt rules; STATIC REVIEW ONLY, no Snort engine test.
# Local SIDs: check for collisions before loading. Use alert, not automatic drop.
# Inspect plaintext HTTP or an authorized decrypted traffic feed after TLS termination.
# Encrypted HTTPS alone does not expose these HTTP buffers. HTTP inspector depth,
# stream reconstruction, decoding, and multipart retention affect coverage.
# SID 1000618: weak caller-supplied label; not exploit or operator identification.
alert tcp any any -> $HOME_NET any (msg:"DC RESEARCH PHP2Shell label in HTTP user agent"; flow:to_server,established; service:http; http_header:field user-agent; content:"php2shell",nocase; sid:1000618; rev:1;)
# SID 1000619: co-occurrence within multipart request body; not a multipart parser
# or proof that the PHP tokens are the upload-progress field value.
# A source-code upload or training text can match. No code execution is established.
alert tcp any any -> $HOME_NET any (msg:"DC RESEARCH multipart upload-progress with PHP command-function tokens"; flow:to_server,established; service:http; http_header:field content-type; content:"multipart/form-data",nocase; http_client_body; content:"PHP_SESSION_UPLOAD_PROGRESS"; content:"<?php",nocase; pcre:"/\b(system|shell_exec|passthru)\s*\(/i"; sid:1000619; rev:1;)

Download php2shell-snort3.rules

Documentation: Snort service matching · Snort HTTP headers · Snort HTTP body buffers · Snort PCRE syntax

kusto

Microsoft Defender XDR: PHP worker child processes

Translate the runtime hunt to Defender Advanced Hunting.

Telemetry
DeviceProcessEvents: InitiatingProcessFileName is the parent basename, FileName is the child basename. DeviceInfo supplies DeviceId and OSPlatform for Linux scoping. Requires applicable onboarded endpoint telemetry.
What it looks for
Seven-day Linux device set and ProcessCreated records; exact PHP-worker basename pattern plus the seven shell/downloader basenames. Returns process IDs, command lines, device, and event identifiers.
Validation
Field names and query operators checked against official documentation; tenant compilation/execution not performed. The shared process predicate has local fixtures, not a Defender-engine test.
Tuning
Expected application shell/download tasks; investigate the service baseline and process command line.
Coverage limits
Missing/stale DeviceInfo, renamed processes, absent ancestry, wrappers, grandchildren, and execution without these child tools. Defender endpoint tables are not a universal HTTP request-body schema.
Read kusto source
// Microsoft Defender XDR Advanced Hunting, not a generic Sentinel table.
// Requires onboarded devices, DeviceInfo, and DeviceProcessEvents telemetry.
// Static schema/logic review only; not compiled or run in a Defender tenant.
// Generic direct-child runtime hunt, not a validated PHP2Shell exploit detector.
let LinuxDevices = DeviceInfo
    | where Timestamp > ago(7d)
    | summarize arg_max(Timestamp, OSPlatform) by DeviceId
    | where OSPlatform == "Linux"
    | project DeviceId;
DeviceProcessEvents
| where Timestamp > ago(7d)
| where DeviceId in (LinuxDevices)
| where ActionType == "ProcessCreated"
| where InitiatingProcessFileName matches regex @"^php-(fpm|cgi)([0-9]+([.][0-9]+)*)?$"
| where FileName in ("sh", "bash", "dash", "ksh", "zsh", "curl", "wget")
| project Timestamp, DeviceId, DeviceName, ReportId, AccountName,
          InitiatingProcessFileName, InitiatingProcessId, InitiatingProcessCommandLine,
          FileName, ProcessId, ProcessCommandLine
| order by Timestamp desc

Download defender-php-worker-child.kql

Documentation: DeviceProcessEvents schema · DeviceInfo schema · Kusto matches regex operator

spl

Splunk CIM: PHP2Shell User-Agent

Translate the HTTP label hunt to Splunk CIM Web data.

Telemetry
CIM Web-tagged events with http_user_agent, src, dest, http_method, url, and status. Add verified index/sourcetype restrictions for your deployment.
What it looks for
Search seven days of tag=web records for the case-insensitive php2shell substring in http_user_agent; retain request context.
Validation
CIM field mapping and SPL search form reviewed against official documentation. Splunk parser/tenant execution was not performed.
Tuning
Authorized testing, research scanners, copied labels, and log replay.
Coverage limits
Absent CIM extraction/tags, unlabelled requests, and body-only behavior. A status code does not prove exploitation.
Read spl source
tag=web earliest=-7d http_user_agent="*php2shell*"
| table _time src dest http_method url status http_user_agent

Download splunk-web-php2shell-label.spl

Documentation: Splunk CIM Web fields

spl

Splunk CIM: PHP worker child processes

Translate the runtime predicate to CIM Endpoint.Processes fields.

Telemetry
CIM process/report tags with parent_process_path and process_path. The process field is the command line. Verify and apply your deployment's Linux os values or restrict to a known Linux process source; the provided query does not apply an OS filter.
What it looks for
Seven-day event search with case-sensitive full-path basename regexes for a direct PHP worker and the seven child tools. Returns host, user, process IDs, paths, and command line.
Validation
Official CIM process tags and fields reviewed; shared local predicate fixtures passed. SPL compilation and tenant execution were not performed.
Tuning
Legitimate process-launch features. Without local Linux scoping, similar Unix process paths may also match.
Coverage limits
Unmapped tags/paths, missing parent telemetry, wrappers, renamed files, grandchildren, and non-child execution. An empty search is not evidence that the service was safe.
Read spl source
tag=process tag=report earliest=-7d
| where match(parent_process_path,"/php-(fpm|cgi)([0-9]+([.][0-9]+)*)?$")
    AND match(process_path,"/(sh|bash|dash|ksh|zsh|curl|wget)$")
| table _time dest user parent_process_path parent_process_id process_path process_id process

Download splunk-php-worker-child.spl

Documentation: Splunk CIM Endpoint fields

logscale

CrowdStrike Falcon LogScale: PHP worker child processes

Translate the runtime hunt to native Falcon process telemetry.

Telemetry
ProcessRollup2 with event_platform=Lin; ParentBaseFileName is the parent basename and ImageFileName is the child path. Requires a Falcon repository retaining these native fields.
What it looks for
Case-sensitive PHP-worker parent and shell/downloader child regexes; returns aid, device name, parent/child IDs, path, and command line. Choose the hunt interval in the time picker.
Validation
Official Falcon/LogScale field and query examples reviewed. Local shared process fixtures are not a LogScale parser test; Falcon tenant execution was not performed.
Tuning
Known PHP service behavior that intentionally launches shells or download utilities.
Coverage limits
Missing or transformed native fields, absent ancestry, renamed tools, wrappers, grandchildren, and the table limit of 1,000 results. This is LogScale syntax, not legacy Falcon Splunk Event Search.
Read logscale source
// Falcon process telemetry in LogScale/Advanced Event Search.
// Not legacy Splunk-based Falcon Event Search syntax.
// Static schema/logic review only; not compiled or run in a Falcon tenant.
#event_simpleName=ProcessRollup2 event_platform=Lin
| ParentBaseFileName=/^php-(fpm|cgi)([0-9]+([.][0-9]+)*)?$/
| ImageFileName=/\/(sh|bash|dash|ksh|zsh|curl|wget)$/
| table([@timestamp, aid, ComputerName, ParentBaseFileName, ParentProcessId,
         ImageFileName, TargetProcessId, CommandLine], limit=1000)

Download falcon-php-worker-child.logscale

Documentation: Falcon process-query examples · Falcon process fields and event tags

sumo

Sumo Logic: Apache access-log User-Agent

Translate the HTTP label hunt using Sumo's documented Apache access parser.

Telemetry
Apache access logs compatible with public/apache/access. The query uses user_agent, src_ip, method, url, and status_code. Adapt the example source selectors to your own logs and set the hunt time range.
What it looks for
Parse the access record, then match the case-insensitive php2shell substring only in user_agent; return request context.
Validation
Parser fields and matches syntax checked against official Sumo documentation. Sumo tenant parsing/execution was not performed.
Tuning
Authorized scanners, copied labels, research, and log replay.
Coverage limits
Missing UA or incompatible log format; the Apache parser does not supply request bodies or EDR ancestry. This translation covers HTTP telemetry; it does not query endpoint process events.
Read sumo source
_sourceName=*access_log* AND _sourceCategory=*apache*
| parse using public/apache/access
| where user_agent matches /(?i).*php2shell.*/
| fields _messageTime, src_ip, method, url, status_code, user_agent

Download sumo-web-php2shell-label.sumo

Documentation: Sumo Apache access parser · Sumo Apache field extraction · Sumo matches operator

spl

Splunk CIM: traffic to the shared download address

Find destination-IP matches across protocols and ports without dropping records with missing direction.

Telemetry
CIM Network Traffic events tagged network and communicate with dest_ip. Scope to relevant network/egress sources. Direction is relative to the reporting device and may be absent or n/a.
What it looks for
Seven-day exact destination-IP match. The outbound filter is commented out; enable it only after validating local direction mapping. Review direction, action, and bytes_out; inbound or unknown-direction rows can match, and a result does not prove /sh was downloaded or run.
Validation
CIM fields and SPL comment syntax checked against official documentation; synthetic fixtures cover the default and optional outbound predicates separately. No Splunk tenant or network replay test.
Tuning
Authorized investigation or testing, shared hosting, address reassignment, and stale indicators. Confirm the time window and ownership context.
Coverage limits
Missing CIM tags/dest_ip extraction, NAT or proxy destinations, rotated addresses, and unlogged traffic. Enabling the optional filter excludes missing direction. The address is an observed submitted destination, not a confirmed C2 role.
Read spl source
tag=network tag=communicate earliest=-7d dest_ip="217.60.102.5"
```Optional narrowing after validating direction mapping; enable the next commented command only when appropriate.```
```| search direction="outbound"```
| table _time src src_ip src_port dest dest_ip dest_port transport action direction dvc bytes_out
| sort 0 - _time

Download splunk-shared-loader-outbound.spl

Documentation: Splunk CIM Network Traffic fields · Splunk search comments

kusto

Microsoft Defender XDR: shared-address network events and process context

Pivot from the shared address to the device, initiating process, parent, and command line.

Telemetry
DeviceNetworkEvents from applicable Defender for Endpoint devices. This table supplies RemoteIP and initiating-process context; it does not provide a universal HTTP User-Agent or response-body field.
What it looks for
Seven-day exact RemoteIP match without filtering platform or action. Inspect ActionType using the portal schema to establish event direction and outcome; failed or inbound events can also be returned.
Validation
Fields reviewed against Microsoft documentation; six inert local destination predicates passed. No Defender tenant compilation or execution.
Tuning
Authorized investigation, security tests, shared infrastructure, and address reassignment. Review the originating process and event outcome.
Coverage limits
Unonboarded devices, unavailable Linux network events, proxies, rotated addresses, missing process fields, and retention gaps. A matching event does not establish a completed download, execution, or malware family.
Read kusto source
// Microsoft Defender XDR Advanced Hunting; requires Defender for Endpoint.
// Network-event pivot with process context; inspect ActionType for outcome and
// direction in the portal schema. This query does not assert a completed download.
// Static schema review and inert local predicate tests, not tenant execution.
DeviceNetworkEvents
| where Timestamp > ago(7d)
| where RemoteIP == "217.60.102.5"
| project Timestamp, DeviceId, DeviceName, ReportId, ActionType,
          LocalIP, LocalPort, RemoteIP, RemotePort, RemoteUrl, Protocol,
          InitiatingProcessFileName, InitiatingProcessId,
          InitiatingProcessCreationTime, InitiatingProcessCommandLine,
          InitiatingProcessParentFileName, InitiatingProcessAccountName
| order by Timestamp desc

Download defender-shared-loader-network.kql

Documentation: DeviceNetworkEvents schema

yaml

Sigma: loader pipeline or SCP remote path

Find the vector-bearing curl/wget-to-sh pattern or the SCP remote operand present in retained command text.

Telemetry
Linux process-creation events with complete CommandLine. The pipeline is usually in a shell -c event; a downloader child may omit it. The SCP branch can match the scp process command line.
What it looks for
Match either a curl/wget pipe to sh -s with apache, ssh, telnet, or php2shell, OR scp with a dlr@HOST:sh remote operand. The SCP account and basename are constrained; the host is unrestricted. SCP token boundaries exclude suffix lookalikes and do not cross command separators, but cannot establish transfer direction; uploads can also match. The pipeline remains a text pattern, not shell parsing.
Validation
Parsed with pySigma 2.0.1 and checked against inert positive/negative strings. Retained incoming command text checks verify pattern coverage, not process execution. No Sigma backend conversion, shell execution, or EDR tenant test.
Tuning
Authorized installers, SCP transfers using the same account/basename, quoted command examples, tests, and log replay. This is text correlation, not shell parsing.
Coverage limits
Truncation, child-only pipeline records, alternate shells, encoding, renamed tools, changed flags/vector names, and alternate SCP user/path/URI/IPv6 forms. Neither branch identifies an operator, CVE, or successful transfer.
Read yaml source
title: Linux Loader Pipeline Or SCP Remote Path With Observed Traits
id: be6d250c-aa01-4b46-8fe3-5fd3e8d2bfb2
status: experimental
description: >-
    Finds command-line text containing curl or wget piped to sh -s with an
    apache, ssh, telnet, or php2shell argument, OR an scp command containing
    the remote operand dlr@HOST:sh. Transfer direction is not established.
    Destinations are deliberately
    unrestricted; correlate with the shared-destination hunt and ancestry.
    This is a text-pattern hunt, not shell parsing or a CVE signature.
author: Deception Check
date: '2026-10-23'
references:
    - https://sigmahq.io/sigma-specification/specification/sigma-rules-specification.html
logsource:
    category: process_creation
    product: linux
    definition: 'Requires the complete process CommandLine, often from the shell -c event.'
detection:
    selection_pipeline:
        CommandLine|re: '(^|[ /"''(;])(?:curl|wget)[ \t]+[^|\r\n]*\|[ \t]*(?:/[^ \t|;]+/)?sh[ \t]+-s[ \t]+(?:--[ \t]+)?(?:apache|ssh|telnet|php2shell)(?:[ \t;&"''()]|$)'
    selection_scp:
        CommandLine|re: '(^|[ /"''(;])scp[ \t]+(?:[^;|&\r\n]*[ \t])?["'']?dlr@[^ \t:;|&"''\r\n]+:sh(?:[ \t"'')]|$)'
    condition: selection_pipeline or selection_scp
fields:
    - Image
    - CommandLine
    - ParentImage
    - ParentCommandLine
    - User
    - ProcessId
    - ParentProcessId
falsepositives:
    - Authorized installation scripts using the same positional arguments
    - Authorized SCP transfers using a remote file named sh and the account dlr
    - Quoted examples, research, test harnesses, and replayed command lines
level: medium

Download sigma-linux-loader-pipe-vector.yml

Documentation: Sigma rule specification

yaml

Sigma: libredtail-http User-Agent

Find the caller-supplied label seen in the linked HTTP traversal requests.

Telemetry
Parsed incoming web-server or reverse-proxy User-Agent mapped to cs-user-agent; the same field mapping as the PHP2Shell label rule.
What it looks for
Case-insensitive libredtail-http substring match within the parsed User-Agent. Return source, method, and path for review.
Validation
Parsed with pySigma 2.0.1; six inert local predicate fixtures passed. No Sigma backend conversion or tenant test.
Tuning
Authorized research, copied headers, security testing, and replay. A label alone cannot establish RedTail family identity.
Coverage limits
Absent or changed User-Agent, incompatible log parsing, and body-only activity. This rule does not cover the later PHP-tagged downloader, whose User-Agent was absent.
Read yaml source
title: Libredtail HTTP User Agent
id: 8463f0de-92ec-45ee-9a28-1d66593e2259
status: experimental
description: >-
    Finds the caller-supplied libredtail-http label seen in the linked HTTP
    traversal requests. Investigate request paths and surrounding activity.
    A copied label does not establish malware family, operator, or execution.
author: Deception Check
date: '2026-10-23'
references:
    - https://sigmahq.io/sigma-specification/specification/sigma-rules-specification.html
logsource:
    category: webserver
    definition: 'Map cs-user-agent to the parsed incoming HTTP User-Agent value.'
detection:
    selection:
        cs-user-agent|contains: 'libredtail-http'
    condition: selection
fields:
    - c-ip
    - cs-method
    - cs-uri-query
    - cs-user-agent
falsepositives:
    - Authorized research, copied labels, security tests, and log replay
level: low

Download sigma-web-libredtail-user-agent.yml

Documentation: Sigma rule specification

yaml

Sigma: Apache starts a shell or downloader

Add a separate, noisier Apache runtime hunt for deployments that may use mod_php.

Telemetry
Linux process-creation events with full Image and direct ParentImage paths. Confirm mod_php separately through inventory; apache2/httpd names alone do not identify a loaded module.
What it looks for
An apache2 or httpd parent directly creates sh, bash, dash, ksh, zsh, curl, or wget. Kept separate from the PHP-worker rule because Apache's legitimate subprocess behavior differs.
Validation
Parsed with pySigma 2.0.1; eight inert local predicate fixtures passed. No backend conversion, EDR deployment, or exploit test.
Tuning
CGI handlers, expected application subprocesses, piped log helpers, and administration. Baseline by server and application before alerting.
Coverage limits
Wrappers, grandchildren, renamed executables, busybox, missing ancestry, and execution without a new child. A result is generic process behavior, not proof of mod_php exploitation.
Read yaml source
title: Linux Apache Starts Shell or Downloader
id: ce2c0053-14c1-40f7-b0a4-28fed13196cd
status: experimental
description: >-
    Finds direct shell or downloader children of apache2 or httpd. Use on
    inventoried Apache servers, including mod_php deployments. A process name
    does not establish that mod_php is loaded. Baseline CGI, application jobs,
    and server administration before alerting; this is a noisy runtime hunt.
author: Deception Check
date: '2026-10-23'
references:
    - https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w
    - https://sigmahq.io/docs/basics/log-sources.html
logsource:
    category: process_creation
    product: linux
    definition: 'Requires full executable paths for Image and direct ParentImage.'
detection:
    selection_parent:
        ParentImage|re: '/(apache2|httpd)$'
    selection_child:
        Image|re: '/(sh|bash|dash|ksh|zsh|curl|wget)$'
    condition: selection_parent and selection_child
fields:
    - Image
    - CommandLine
    - ParentImage
    - ParentCommandLine
    - User
    - ProcessId
    - ParentProcessId
falsepositives:
    - CGI handlers, authorized application subprocesses, and maintenance tasks
    - Apache piped logging or other configured server helpers invoking a shell
level: low

Download sigma-linux-apache-child.yml

Documentation: PHP OOM advisory · Sigma log-source guidance

spl

Splunk: PHP memory errors, timeouts, and worker exits

Review PHP memory-limit/allocator messages, timeouts, FPM signal-11 or code-255 exits, and Apache child segmentation faults.

Telemetry
Raw PHP/FPM/Apache error-log records with _raw, _time, host, and source. The source globs are examples; replace them with verified local sources and include journald or container logs where applicable.
What it looks for
Classifies six message patterns into distinct families and counts messages in five-minute bins per host and source. Code 255 is generic, not an OOM or request-startup diagnosis. Compare bins with normal behavior and nearby requests; no universal alert threshold is applied.
Validation
SPL syntax and message forms checked against PHP/Apache source, with synthetic predicate tests. Allocated-form OOM text is verified in PHP 7.4.33 (without bytes) and 8.4.26 (with bytes); both forms are covered. These are source-supported fixtures, not observed runtime messages from this research; real-log parsing, source selection, binning, and aggregation need tenant validation.
Tuning
Application bugs, expensive legitimate requests, memory pressure, extension faults, normal load changes, and duplicated log forwarding.
Coverage limits
Disabled, truncated, or differently formatted logs; uncollected container/kernel events; bare OOM and newer hard-limit wording; failures before logging. Versions/builds differ. Counts are messages, not distinct crashes; no CVE or exploitation verdict.
Read spl source
earliest=-7d (source="*php*error*log*" OR source="*php*fpm*log*" OR source="*apache*error*log*" OR source="*httpd*error*log*")
| eval error_family=case(
    match(_raw,"(?i)allowed memory size.*exhausted"),"php_memory_limit",
    match(_raw,"(?i)out of memory [(]allocated [0-9]+(?: bytes)?[)]"),"php_allocator_oom",
    match(_raw,"(?i)child [0-9]+ exited on signal 11(?:[^0-9]|$)"),"fpm_sigsegv",
    match(_raw,"(?i)child [0-9]+ exited with code 255(?:[^0-9]|$)"),"fpm_exit_255",
    match(_raw,"(?i)child pid [0-9]+ exit signal (?:Segmentation fault|11)(?:[^0-9]|$)"),"apache_sigsegv",
    match(_raw,"(?i)maximum execution time.*exceeded"),"php_execution_timeout")
| where isnotnull(error_family)
| bin _time span=5m
| stats count AS event_count values(error_family) AS error_families by _time host source
| sort 0 - event_count

Download splunk-php-worker-errors.spl

Documentation: PHP stale-state advisory · PHP-FPM configuration · FPM child-exit logging source · Apache child-exit logging source · Splunk comparison functions · Splunk bin command · PHP 7.4.33 allocator source · PHP 8.4.26 allocator source · PHP 8.4.27 FPM startup source

Detection catalog and validation notes

Method and limits

This is a retained-log case study, not fleet-wide prevalence. We selected incoming request fields with explicit PHP2Shell identifiers or upload-progress markers, removed exact duplicate request lines, excluded collector mirrors, and matched sources either directly or through a connection with the same sensor identity and session. That broad selection includes the separate WordPress pair and the sender-labelled downloader input. Context requests and sessions are counted separately. The technical guide describes the selection and counting methods.

Sensor-recorded timestamps are not a historical clock attestation. Source IPs and tool labels do not identify an operator. Generated decoy responses and captured input establish receipt, not runtime execution or compromise. The retained selection can miss unlabelled variants or activity outside its coverage. We did not run exploit code or fetch submitted destinations. Public exports exclude targets, raw cookies, session IDs, sensor names, private paths, and internal review records.